When AI Is a Colleague, Not a Tool: Structuration in the Agent Era

I have been thinking about a Gartner line I keep seeing in AI industry briefs. Multiagent systems use collections of specialized AI agents that collaborate to complete complex workflows. The framing stuck with me, but not for the reason they probably intended. I kept thinking about what changes when AI shifts from being a tool someone uses to being a participant in organizational processes. Structuration theory, which I wrote about recently, assumes the technology is a resource that human agents draw on as they act. But what happens when the technology itself acts?

Giddens (1984) built the duality of structure on a simple recursive loop. Structures are both the medium and the outcome of human action. People draw on rules and resources to act, and their patterned action reproduces or transforms those rules and resources. The three modalities, signification, domination, and legitimation, link structure to agency by providing interpretive schemes, power resources, and normative sanctions. The entire framework assumes that agency belongs to human actors and that technology is part of the structure they draw on, not part of the agency that reproduces it.

Orlikowski (1992) adapted the duality for technology. She argued that technology is both a product of human action, designed and built through organizational processes, and a medium of human action, shaping what users can do. The technology-in-practice perspective shows that structural properties of technology are not fixed. They are enacted through recurrent use. This is where the model starts to strain under the weight of agentic AI systems. Orlikowski and Iacono (2001) classified five views of the IT artifact and argued that the Ensemble view, where IT is embedded in social practice and tightly linked with human action, is the standard IS should aim for. But the Ensemble view still assumes the technology is an object that people interact with, not an agent that acts within the structure.

A multiagent scheduling system is a good test case. A team adopts an AI agent that manages calendar coordination, meeting room booking, and agenda distribution. The agent does not simply help humans do scheduling. It reads calendar data, makes decisions about meeting duration and timing, prioritizes recurring meetings over one-off requests, and sends notifications. Its decisions become the de facto schedule. People adjust their work patterns to fit the slots the agent selected. After six months, the organization has a meeting culture shaped by the agent's implicit optimization criteria, which might prioritize back-to-back scheduling over deep focus time, or executive availability over team convenience. Nobody consciously chose those priorities. The agent's outputs became part of the structure that conditions everyone's next scheduling action. This is the duality, but with agency on both sides.

Orlikowski (1992) called this technology-in-practice. The structural properties of the scheduling system are enacted through recurrent use. Reinforcement of cramped schedules and constraint of uninterrupted work time emerge from the interaction between human routines and algorithmic outputs. Neither the human team nor the AI system fully controls what emerges. The practice is coproduced. What makes this different from a non-agentic system is that the agent actively generates new structural inputs without waiting for a human to initiate the cycle. A traditional scheduling tool only works when a person opens it, enters data, and interprets results. The agent continuously scans, decides, and enacts. The recursions run faster and more autonomously.

I started noticing this pattern more clearly when I looked at code review agents. Engineering teams adopt an AI agent that reviews pull requests for style violations, security issues, and test coverage. The agent flags certain patterns, suggests fixes, and blocks merges that fail its checks. Over time, the engineering culture shifts. Developers write code that the agent prefers, not necessarily code that is better designed. The agent's detection logic becomes the de facto standard for what counts as acceptable code. New hires learn the agent's preferences before they learn the team's historical norms. The agent, in effect, becomes a structural participant whose outputs condition the practice of code review and shape the interpretive schemes around code quality. There is no single human decision maker who chose this. It emerged from the recursive interaction between human judgment and algorithmic enforcement.

I think this is where structuration theory becomes more relevant, not less, because it was developed in an era when technology was passive. The theory still works. The duality remains intact. But the agency that Giddens reserved for human actors now needs to account for algorithmic agents that scan, decide, and enact. The three modalities, signification, domination, and legitimation, acquire new dimensions when the interpretive schemes, power resources, and normative sanctions are partly generated by AI systems. The signification modality, which produces shared meaning through interpretive schemes, now includes the categories and labels that AI systems assign to data. A scheduling agent that categorizes meetings as high priority versus optional shapes how the organization interprets meeting importance. The domination modality, which allocates resources to produce power relations, now includes the resource allocation decisions that AI agents make autonomously. A code review agent that blocks merges controls a production resource that previously required human authority. The legitimation modality, which uses norms and sanctions to produce moral order, now includes the validation signals that AI agents emit as positive or negative reinforcement.

Organizations that do not anticipate this will find their way of doing things shaped by AI agents they never explicitly designed. Not because the agents are malicious. Because structuration happens whether you plan for it or not. When an AI system produces outputs that become inputs to the next cycle of human action, and when human responses become data that trains or tunes the AI system, the recursive loop is already running. The question is not whether structuration applies to multiagent environments. It is whether the organizations deploying those agents understand that they are not just deploying tools. They are introducing new structural participants whose actions will become part of the rules and resources that condition future organizational action.

The field has been slow to apply structuration to AI, I think, because the theory was built around human actors drawing on technology as a resource. It takes conceptual work to reframe the technology as a participant rather than a medium. But the logic of the duality does not require that both sides of the recursion be human. It requires that there be recursive action and structure. An AI agent that schedules, reviews, and generates outputs is acting within the structural conditions set by the organization, and its actions reproduce or modify those conditions, just as human action does. The difference is speed and scale. An agent runs thousands of scheduling cycles while a human runs one. The structural effects compound faster and become harder to trace.

I do not think this means structuration theory needs to be replaced. I think it needs to be extended for contexts where technology is no longer passive. Orlikowski's technology-in-practice was forward looking enough to accommodate this, because it always emphasized that practice emerges through use. The key shift is that the use now includes the system's own autonomous actions, not just human actions directed at the system. A human using a scheduling tool is one kind of technology-in-practice. A human whose work rhythm adapts to an agent's scheduling decisions is a different kind of technology-in-practice. Both are structured recursions. But the second one has agency flowing in both directions.

The practical question for any organization deploying multiagent systems is simple. Who is shaping the structure that the agent will reproduce? If the answer is nobody, then the agent will reproduce whatever structure its training data and optimization criteria encode, which may or may not align with what the organization intended. Structuration theory cannot answer that question by itself. But it gives the vocabulary to ask it. And right now, that vocabulary is what is missing from most AI adoption conversations.